Sixth International Conference on Spoken Language Processing
Utterance verification based on N-Best HMM scores has been widely used in ASR system. There are a number of ways to calculate a measurement score for verification from N-Best scores. Most of proposed methods are based on the framework of the hypothesis testing. This has lead to use the second best score or an overall average of available N-Best scores for normalisation. In this study we examine N-Best UV approach from a competition-based measurement framework. With this framework different competitive measurements can be derived from a sequence of sorted likelihood ratios (SLR). The evaluation results demonstrate that OOV performance can be improved by using some selective components in SLR. In our experiments by using the first four components OOV rejection errors can be reduced about 30% in comparison with the baseline results.
Bibliographic reference. Gu, Yong / Thomas, Trevor (2000): "Competition-based score analysis for utterance verification in name recognition", In ICSLP-2000, vol.4, 214-217.